HARMONY: New Framework for Heterogeneous Split Federated Learning
HARMONY represents the inaugural hybrid split federated learning (SFL) framework tailored for various client architectures. It tackles the issue of representation skew, which occurs when features from tailored client-side extractors do not correspond in the shared space, negatively impacting the server model's effectiveness in out-of-distribution (OOD) predictions. The framework innovatively adjusts meta-learning to replicate diverse extractors across different parameters and architectures, facilitating personalization without sacrificing generalization. This initiative focuses on mobile devices characterized by varying resource limitations and non-IID data distributions, striking a balance between accuracy and cost via early exit and fallback inference strategies.
Key facts
- HARMONY is the first hybrid SFL framework to support heterogeneous client architectures.
- It mitigates representation skew in hybrid split federated learning.
- The framework modifies meta-learning to simulate diverse extractors.
- It addresses non-IID data class distributions and resource constraints on mobile devices.
- Hybrid SFL couples personalized client-side front ends with a generalized server-side backend.
- Representation skew causes degradation in server model for OOD prediction.
- The work is published on arXiv with identifier 2605.07211.
- The announcement type is cross.
Entities
Institutions
- arXiv